pytorch實現(xiàn)focal loss的兩種方式小結(jié)

import torch
import torch.nn.functional as F
import numpy as np
from torch.autograd import Variable
'''
pytorch實現(xiàn)focal loss的兩種方式(現(xiàn)在討論的是基于分割任務)
在計算損失函數(shù)的過程中考慮到類別不平衡的問題，假設加上背景類別共有6個類別
'''
def compute_class_weights(histogram):
  classWeights = np.ones(6, dtype=np.float32)
  normHist = histogram / np.sum(histogram)
  for i in range(6):
    classWeights[i] = 1 / (np.log(1.10 + normHist[i]))
  return classWeights
def focal_loss_my(input,target):
  '''
  :param input: shape [batch_size,num_classes,H,W] 僅僅經(jīng)過卷積操作后的輸出，并沒有經(jīng)過任何激活函數(shù)的作用
  :param target: shape [batch_size,H,W]
  :return:
  '''
  n, c, h, w = input.size()

  target = target.long()
  input = input.transpose(1, 2).transpose(2, 3).contiguous().view(-1, c)
  target = target.contiguous().view(-1)

  number_0 = torch.sum(target == 0).item()
  number_1 = torch.sum(target == 1).item()
  number_2 = torch.sum(target == 2).item()
  number_3 = torch.sum(target == 3).item()
  number_4 = torch.sum(target == 4).item()
  number_5 = torch.sum(target == 5).item()

  frequency = torch.tensor((number_0, number_1, number_2, number_3, number_4, number_5), dtype=torch.float32)
  frequency = frequency.numpy()
  classWeights = compute_class_weights(frequency)
  '''
  根據(jù)當前給出的ground truth label計算出每個類別所占據(jù)的權(quán)重
  '''

  # weights=torch.from_numpy(classWeights).float().cuda()
  weights = torch.from_numpy(classWeights).float()
  focal_frequency = F.nll_loss(F.softmax(input, dim=1), target, reduction='none')
  '''
  上面一篇博文講過
  F.nll_loss(torch.log(F.softmax(inputs, dim=1)，target)的函數(shù)功能與F.cross_entropy相同
  可見F.nll_loss中實現(xiàn)了對于target的one-hot encoding編碼功能，將其編碼成與input shape相同的tensor
  然后與前面那一項（即F.nll_loss輸入的第一項）進行 element-wise production
  相當于取出了 log(p_gt)即當前樣本點被分類為正確類別的概率
  現(xiàn)在去掉取log的操作，相當于 focal_frequency shape [num_samples]
  即取出ground truth類別的概率數(shù)值，并取了負號
  '''

  focal_frequency += 1.0#shape [num_samples] 1-P（gt_classes）

  focal_frequency = torch.pow(focal_frequency, 2) # torch.Size([75])
  focal_frequency = focal_frequency.repeat(c, 1)
  '''
  進行repeat操作后，focal_frequency shape [num_classes,num_samples]
  '''
  focal_frequency = focal_frequency.transpose(1, 0)
  loss = F.nll_loss(focal_frequency * (torch.log(F.softmax(input, dim=1))), target, weight=None,
           reduction='elementwise_mean')
  return loss


def focal_loss_zhihu(input, target):
  '''
  :param input: 使用知乎上面大神給出的方案 https://zhuanlan.zhihu.com/p/28527749
  :param target:
  :return:
  '''
  n, c, h, w = input.size()

  target = target.long()
  inputs = input.transpose(1, 2).transpose(2, 3).contiguous().view(-1, c)
  target = target.contiguous().view(-1)

  N = inputs.size(0)
  C = inputs.size(1)

  number_0 = torch.sum(target == 0).item()
  number_1 = torch.sum(target == 1).item()
  number_2 = torch.sum(target == 2).item()
  number_3 = torch.sum(target == 3).item()
  number_4 = torch.sum(target == 4).item()
  number_5 = torch.sum(target == 5).item()

  frequency = torch.tensor((number_0, number_1, number_2, number_3, number_4, number_5), dtype=torch.float32)
  frequency = frequency.numpy()
  classWeights = compute_class_weights(frequency)

  weights = torch.from_numpy(classWeights).float()
  weights=weights[target.view(-1)]#這行代碼非常重要

  gamma = 2

  P = F.softmax(inputs, dim=1)#shape [num_samples,num_classes]

  class_mask = inputs.data.new(N, C).fill_(0)
  class_mask = Variable(class_mask)
  ids = target.view(-1, 1)
  class_mask.scatter_(1, ids.data, 1.)#shape [num_samples,num_classes] one-hot encoding

  probs = (P * class_mask).sum(1).view(-1, 1)#shape [num_samples,]
  log_p = probs.log()

  print('in calculating batch_loss',weights.shape,probs.shape,log_p.shape)

  # batch_loss = -weights * (torch.pow((1 - probs), gamma)) * log_p
  batch_loss = -(torch.pow((1 - probs), gamma)) * log_p

  print(batch_loss.shape)

  loss = batch_loss.mean()
  return loss

if __name__=='__main__':
  pred=torch.rand((2,6,5,5))
  y=torch.from_numpy(np.random.randint(0,6,(2,5,5)))
  loss1=focal_loss_my(pred,y)
  loss2=focal_loss_zhihu(pred,y)

  print('loss1',loss1)
  print('loss2', loss2)
'''
in calculating batch_loss torch.Size([50]) torch.Size([50, 1]) torch.Size([50, 1])
torch.Size([50, 1])
loss1 tensor(1.3166)
loss2 tensor(1.3166)
'''